A spatial domain Optimal Trade-off Maximum Average Correlation Height (SPOT-MACH) filter is proposed in this thesis. The proposed technique uses a pre-defined fixed size kernel rather than using estimation techniques. The spatial domain implementation of OT-MACH offers the advantage that it does not have shift invariance imposed on it as the kernel can be modified depending upon its position within the input image. This allows normalization of the kernel and allows inclusion of a space domain non-linearity to improve performance. The proposed SPOT-MACH filter can be used to maximize the height of the correlation peak in the presence of distortions of the training object and provide resistance to background clutter. One of the major characteristics of the SPOT-MACH filter is that it can be tuned to maximize the height and sharpness of the correlation peak by using trade-offs between distortion tolerance, peak sharpness and the ability to suppress clutter noise. A number of non-parametric local regression techniques offer a simplified approach to pattern recognition problems which employ linear filtering using low pass filters designed using moving window local approximations. In most of these cases the algorithms search for a region of interest near the point of estimation for various prevailing conditions which fit the required criteria. These estimates are calculated for a defined window size which is determined as being the largest area within which the estimators do not widely vary from the criteria. The only drawback in this approach is that the window size is directly proportional to the required computational resources and would adversely affect the performance of the system if the moving window size is not proportionate to the resources. The proposed filter employs an optimization technique using low-pass filtering to highlight the potential region of interests in the image and then restricts the movement of the kernel to these regions to allow target identification and to use less computational resources. Also another optimization technique is also proposed which is based on an entropy filter which measures the degree of randomness between two changing scenes and would return the area where change has occurred i.e. the target object might be present. This approach gives a more accurate region of interest than the low-pass filtering approach. Apart from the software based optimization approaches two hardware based enhancement techniques have also been proposed in this thesis. One of the approaches employs Field Programmable Gate Array (FPGA) to perform correlation process employing the inbuilt multipliers and look up tables and the other one uses Graphical Processing Unit (GPU) to do parallel processing of the input scene. Also in this thesis a detailed analysis of SPOT-MACH has been carried out by comparing with popular feature based techniques like Scale Invariant Feature Transform (SIFT) and a comparison matrix has been created. The proposed filter uses a two-staged approach using speed optimizations and then detection of targets from input scenes. Both visible and Forward Looking Infrared (FLIR) imagery data sets have been used to test the performance of filter.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:582063 |
Date | January 2013 |
Creators | Gardezi, Akber Abid |
Publisher | University of Sussex |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | http://sro.sussex.ac.uk/id/eprint/47055/ |
Page generated in 0.0022 seconds